Pushing the Envelope: Stretching the Limits of Generative

António Leitão Luís Santos INESC-­‐ID/Instituto Superior Técnico, Portugal University of California, Berkeley, USA [email protected] [email protected]

Rita Fernandes Instituto Superior Técnico, Portugal [email protected]

Abstract Design is characterized by change. Nowadays, addressing change in design requires enormous efforts, particularly, when the is exploring different solutions or when it is necessary to adapt the design to evolving constraints. This paper discusses the potential of Generative Design to help handling change, namely in the generation of several alternative design solutions. We propose a programming-­‐based approach that, although requiring an initial investment, dramatically reduces the efforts of design modification. We evaluate the approach in a case study and we show that it is cost-­‐effective in the development of an architectural design.

Keywords: Design process; Change; Modeling; Generative d esign; Programming .

Introduction Case Study Generative Design (GD) (McCormack, 2004) helps overcome these In order to answer the previous question, we made an limitations. GD can be summarily described as form creation experiment: we formalized the design of a building in a GD through (Terdizis, 2003). One of the main benefits of program which supports as many changes as possible. We then 235 GD is the fast and effortless generation of a wide range of tested the resilience of our formalization by simulating several solutions, exploring different design approaches and implementing realistic changes and measuring their impact, particularly, in the the continuously changing constraints and requirements which time and effort required for their implementation compared to characterize the design process. Thus, it supports the architects in the conventional use of CAD . tools Our case study was focused on the exploration of a larger solution as well as in the a large and complex commercial building that is under conceptualization of innovative solutions through the interaction construction: MVRDV’s Market Hall (Boranyak, 2010). This building between design constraints and (Kilian, parameters 2006). is characterized by an unconventional bent shape, which is closed by two glass facades. Because it constraints both the building Writing a GD program that describes an elements and the interior spaces, the shape of the building was requires a large initial effort, and might be less -­‐ cost effective than the starting point for our modeling approach. However, instead of the typical digital modeling approach. However, this initial cost capturing the exact geometry of the building, we captured the can be quickly recovered when it becomes necessary to underlying ideas of the design, and formalized them in a incorporate changes in . the design This happens because the corresponding GD program. design is unambiguously represented in a parameterized GD program and a significant number of changes only require The modeling process we used can be divided into four phases: (1) different values for its parameters. These parameters might formalization of the building shape, (2) modeling of building represent numerical values, geometric shapes, mathematical elements (walls, slabs, etc.), (3) openings and frames, and finally functions, or even subprograms. The important idea is that by (4) detailed elements (posts, the elements of a glass facade, etc.). changing the current values of the parameters or by including In the first phase, we defined the underlying geometry that additional parameters, the designer can quickly generate a captures the overall shape of the building. This geometry was then different model that expresses the changes in the design. used as input for the remaining phases of the modeling process. Obviously, not all changes are trivial to implement and some might The decomposition of the process into four phases, each one require additional efforts to adapt the GD program. Thus, the increasing the detail of the model, led to the existence of research question which can be posed is: how far can we go with dependencies between elements. Their sequence ensures that the this approach? geometry produced in each stage fits in the one produced in previous phases, allowing automatic change propagation. This

Parametric Modeling means that a change in, e.g., the building form, causes a between elements were kept and adapted to a different scale. correspondent change in the window openings, frames, shape of Similarly to the first example, these results were produced by the walls and slabs, etc. changing a small set of parameters, a task that was implemented in minutes. Compared to the previous example, these changes are It is precisely these changes that are visible in Figure 1. The variant even more difficult to handle in a conventional approach because building (below) was generated by simply changing the building it is not possible to reuse any element from one model to the next. form. The remaining parameters were kept identical to those used to generate the Market Hall (top). The instanced model reveals a In the first two scenarios we modified the actual values of large number of differences in the elements of , the building parameters but it is also important to consider changes that also including their geometry, position, direction, and number of sub-­‐ require modifications in the e impl mented algorithms. For elements. example, Figure 3 illustrates changes on the geometry of the fencing posts.

To implement this change we generalized the parts of the program that dealt with fencing posts so that, instead of always using a rectangular section (identical to those of the Market Hall, visible in the top), they now accept an additional parameter for a procedure that produces the desired shape. Figure 3 shows this feature by replacing the default rectangular with section a circular one.

236

Figure 1: Top: MVRDV’s Market Hall as generated from our GD program. Below: A variant building generated by changing the overall shape of the building.

Usually, changes in the overall shape of a building are very difficult to accommodate using conventional modeling approaches. However, using our GD-­‐based approach, we implemented the change in minutes.

In the previous example we simulated just a refinement in the

building shape, without changing its general dimensions. In the next example we simulate a more extensive change, not only in the shape, but also in the positions and dimensions of specific elements of the building. In Figure 2 we wo show t outcomes resulting from changing the parameters that control the form, the Figure 2: Two buildings generated by introducing several changes in the number and position of the slabs and walls, the number of shape and in the number and position of slabs, walls, windows, and other windows, their position and dimensions, etc. The relationships elements.

SIGraDi 2013

Figure 3: Top: Posts with rectangular sections. Below: Posts with circular Figure 4: A parallelepiped building generated by changing the sections. formalization of the overall shape of the Market Hall. 237 The modifications de we ma to our GD program were completed in In Table 1 we present the results of our inquiry. The analysis shows less than an hour, a very short amount of time for a task that, in a that the expected times of the conventional approach are, without conventional approach, requires changing a large number of exception, much longer than the actual times of our approach. As elements and, moreover, must be repeated each time we try a a result, the inquiry corroborated our belief that GD helps different alternative. On the contrary, n i our approach, the designers handling change. generalization effort only needs to be done once, allowing the designer to effortlessly test many different alternatives. Table 1: Results of the inquiry: comparison of the time needed to implement the changes between the conventional and -­‐ the programming based approach. Finally, we simulated a more drastic scenario: changing the building from a bent and longitudinal shape to a parallelepiped Time (hours) Conventional Programming-­‐ shape (Figure . 4) In this case, we were forced to implement Change approach based approach additional algorithms that produce the specific geometry required, Building overall shape 70 < 1/6 and change some of the program parameters. The important (Figure 1) point, however, is that all these changes were made in Building overall shape, approximately half an hour. If an equivalent reuse is intended in a dimensions and 92 < 1/4 conventional approach, the designer might reuse some elements positions of specific from other models, such as the frames of the lateral facades, but elements (Figure 2) only when their overall dimensions . are identical This is such a Shape of specific severe restriction that, in practice, designers rebuild their models elements -­‐ fencing 18 < 1 from scratch. posts (Figure 3) Formalization of the Evaluation building overall shape 42 < ½ (Figure 4) To validate the results of our experiment we did an inquiry, asking 10 experts in several CAD tools to determine the expected time to During this experiment we noticed that when a change in the implement any hese of t changes using their conventional overall shape was required, all the participants in the inquiry approaches. would throw away their previous models and start the modeling

Parametric Modeling process from scratch. At most, they would just try to reuse and Although a larger initial effort is needed to formalize a design, our adjust specific elements with standard geometries, e.g., experiment shows that this effort is quickly recovered. Our rectangular frames or revolving doors. This shows that, despite the evaluation proved that this approach is sufficiently flexible, not underlying logic inherent to a specific design and the application of only to accommodate expected changes, but also changes that the same modeling process to change its features, the major were not planned. This conclusion can be extended to the modeling task would always be repeated. This causes a architectural practice, where changes requently f arise without tremendous waste of time and effort that was already spent in being anticipated. previous modeling activities. An important lesson from this work is that the developed program Although our proposed approach shows large gains in handling should attempt to generalize the design as a mean to simplify the change, these gains depend on the development of a GD program generation of different instances from an initial . model This s i that, obviously, requires time and effort. As a result, it is crucial to possible due he to t creation of parameter and variable this research to also evaluate this effort. To compare the time interdependencies that automatically propagate changes between required to develop this program with the conventional modeling building elements. The proposed approach is aligned with the task we made a second inquiry to the same articipants p , asking requirements of the design process, by allowing an effortless them about the time they would need to model the Market Hall introduction of changes and also the exploration of different building. As we predicted, the results, shown in Table 2, proved design solutions, thus assisting the designers in decision-­‐making that the initial development cost using our approach is larger than activities. Our evaluation shows that a programming approach can using the conventional one. However, when we analyze the also support large scale and can easily incorporate mass-­‐ cumulative time to implement the initial model and handling the customization strategies (Duarte, 2005), in which the effort changes, we verify that the initial investment in the GD program required to produce one program is recovered by its use to model was highly rewarded. several related buildings.

Table 2: Results of the inquiry: comparison of the time needed to produce Finally, we should mention an important advantage of the the model / develop the GD program. programming-­‐based approach: any inconsistencies or errors in the Time (hours) design quickly show up as bugs in the program. This allows early Conventional Programming-­‐ discovery of problems that are more costly to solve using 238 approach based approach Production of the conventional approaches. 131 160 model / program Acknowledgments Conclusions This work was partially supported by nal Portuguese natio funds We designed this experiment to determine how far we can go through FCT under contract Pest-­‐OE/EEI/LA0021/2013, by the using a pure programming approach to architectural design. The Rosetta project under contract PTDC/ATP-­‐AQI/5224/2012, by the results of the experiment allow us to conclude that it is possible to project PTDC/AUR-­‐AQI/103434/2008 and by the PhD scholarship go very far indeed. SFRH/BD/91939/2012.

Given that this was an experiment, we decided to stop it as soon References as we had collected enough information, but it was clear to us that we could go much further. Although more experiments are Boranyak, S. (2010). Archetype. Civil , ASCE, 80(2), 76-­‐79. needed (e.g. implementing this programming-­‐based approach in Duarte, J. P. (2005). Towards the mass customization of housing: the an office during a real architectural design process), grammar of Siza’s houses at Malagueira. Environment and planning we believe that it is not only possible but actually cost-­‐effective to B: Planning and Design, 32(3), 347-­‐380.

develop designs with the aid of a pure programming approach. Kalay, . Y (2004). Architecture's New Media: Principles, Theories, and Methods of Computer-­‐Aided Design. Cambridge, Massachusetts: The Designs that can be formalized by explicit and implicit patterns MIT Press. and rules, such as the Market Hall, are prime candidates for Kilian, A. (2006). Design innovation through constraint modeling. applying this approach. Our experiment proved that a International Journal of Architectural Computing, 1(4), 87-­‐105. programming modeling method can be integrated in the design McCormack, J., Dorin, A., & Innocent, T. (2004). Generative design: a process right after the first crystallization of the architectural paradigm for . Proceedings of Futureground, Design concept. This approach may be also useful during the , Melbourne. conceptualization phase, allowing the quick production of simple Terdizis, K. (2003). Expressive Form: A Conceptual Approach to models that explore different conceptual alternatives. Computational Design. London and New Spon York: Press.

SIGraDi 2013